Conference PaperPDF Available

Survey of Fall detection techniques based on computer vision

Authors:

Abstract

Falls affect tens of millions of older people throughout the world, approximately 28 people over the age of 65 falls. Moreover, falling accidents may lead to a set of symptoms after fall, which includes loss of autonomy, confusion, increased dependence, analyzation, and depression. We can say that falls are the primary cause of injury related to death for elderly and the second leading cause of injury-related death for all ages. Falling accidents, the rapid movement from standing or sitting to a lying position can be classified into four pre-fall, critical fall, post-fall, and recovery phases. In this paper, we have reviewed and compared different fall detection techniques based on computer vision in four phases of fall. Examples of techniques founded are Bayesian Network, Centroid, Eigen Gaussian mixture model: GMM, Hidden Markov model: HMM, Neural Network, Support vector machine: SVM, and velocity. Most of the fall detection techniques are focused on critical phase and post, which mean that the falling person had fallen already and may be injured. Sensor technologies and multi-camera systems were used to improve the results in the pre-early critical fall phase by combining information from several cameras and sensors to make a decision. Microphone array was also used along with a vision system to determine the direction of the audio signal from multiple sources so that it is possible to track many people at the same time.
Survey of Fall detection techniques based on computer vision
Nuth
School of Informatics, Walailak University, Nakornsi Thammarat, Thailand
auisuke@gmail.com, poon
Abstract
Falls affect tens of millions
of elderly people
throughout the world, approximately 28
of people over the age of 65 fall
Moreover, falls may lead to set of symptoms
after fall, which includes loss of autonomy,
confusion, increased dependence,
zation and depression.
We can say that falls are
the primary cause of injury related to death for
elderly and the second leading cause of injury
related death for all ages.
Falling, the rapid
movement from standi
position
can be classified in to four p
pre-fall, critical fall, post-
fall and recovery
phases.
In this paper, we have reviewed and
compared different fall detection techniques
based on computer vision in four phases of
fall.
Examples of techniques founded are
Bayesian Network, Centroi
d, Eigen
Gaussian mixture model: GMM, Hidden
Markov model: HMM, Neural Network,
Support vector machine: SVM and velocity.
Most of the fall detection techniques are
focused on critical phase and post
which mean that the falling person had
fallen already and may be injured.
sensor technologies and multi-
camera systems
were used to improve the results in pre
early critical fall phase by combining
information from several cameras and sensors
to make decision. Microphone
array were also
used along with vision system to determine the
direction of the audio signal from multiple
sources so that it is possible to track many
people at the same time.
Keywords:
Fall detection; Fall activities;
Elderly
1. Introduction
“Population ageing is a triumph of humanity
but also a chal
lenge to society” had been
declared by
World Health Organization
[28].
Statistics in Global brief for World
Survey of Fall detection techniques based on computer vision
Otanasap and Poonpong Boonbrahm
School of Informatics, Walailak University, Nakornsi Thammarat, Thailand
auisuke@gmail.com, poon
pong@gmail.com
of elderly people
throughout the world, approximately 28
-35%
of people over the age of 65 fall
s each year.
Moreover, falls may lead to set of symptoms
after fall, which includes loss of autonomy,
confusion, increased dependence,
immobili-
We can say that falls are
the primary cause of injury related to death for
elderly and the second leading cause of injury
Falling, the rapid
ng or sitting to lying
can be classified in to four p
hases as
fall and recovery
In this paper, we have reviewed and
compared different fall detection techniques
based on computer vision in four phases of
Examples of techniques founded are
d, Eigen
-space,
Gaussian mixture model: GMM, Hidden
Markov model: HMM, Neural Network,
Support vector machine: SVM and velocity.
Most of the fall detection techniques are
focused on critical phase and post
-fall phase
which mean that the falling person had
been
fallen already and may be injured.
Some new
camera systems
were used to improve the results in pre
-fall and
early critical fall phase by combining
information from several cameras and sensors
array were also
used along with vision system to determine the
direction of the audio signal from multiple
sources so that it is possible to track many
Fall detection; Fall activities;
“Population ageing is a triumph of humanity
lenge to society” had been
World Health Organization
(WHO)
Statistics in Global brief for World
Health Day 2012 [27]
show that approximately
28-
35% of people over the age of 65 fall each
year, and this proportion increases to 32
for those aged more than 70 years.
falls may lead to set of symptoms after fall,
which includes loss of autonomy, confusion,
in
creased dependence, immobilization and
depression.
That means falls are the primary
reason of injury related death for seniors and
the second leading cause of injury related
death for all ages.
WHO
“fall” as an event which results in a person
coming to rest inadvertently on the ground or
floor or other lower level.
raises the interest of researchers, particularly
preventive detection of the fall because there
are inapprop
riately defined processes and can
be approached using various methods.
reduce the risk of falling, the early detection of
fall in pre-
fall or critical fall phase therefore
advances the interest of researchers.
research has been done in this
algorithms and systems of preventive fall
detection.
The goals of this study were to
classify the computer vision based approaches
to prevented fall and to comparing the results
of these studies.
2. Fall detection definition
Fig. 1:
Four phases of a fall event.
According to daily life motions, Noury et al.
[15]
had described a fall event in four phases
as shown in Fig. 1.
Firstly, the pre
sudden movements directed to the ground as
sitting or crouching down occ
second phase is critical phase that very short
period and most important of fall event.
Survey of Fall detection techniques based on computer vision
School of Informatics, Walailak University, Nakornsi Thammarat, Thailand
show that approximately
35% of people over the age of 65 fall each
year, and this proportion increases to 32
-42%
for those aged more than 70 years.
Moreover,
falls may lead to set of symptoms after fall,
which includes loss of autonomy, confusion,
creased dependence, immobilization and
That means falls are the primary
reason of injury related death for seniors and
the second leading cause of injury related
WHO
[28] had defined
“fall” as an event which results in a person
coming to rest inadvertently on the ground or
floor or other lower level.
The fall therefore
raises the interest of researchers, particularly
preventive detection of the fall because there
riately defined processes and can
be approached using various methods.
To
reduce the risk of falling, the early detection of
fall or critical fall phase therefore
advances the interest of researchers.
Several of
research has been done in this
area to develop
algorithms and systems of preventive fall
The goals of this study were to
classify the computer vision based approaches
to prevented fall and to comparing the results
2. Fall detection definition
Four phases of a fall event.
[15, 19]
According to daily life motions, Noury et al.
had described a fall event in four phases
Firstly, the pre
-fall phase is
sudden movements directed to the ground as
sitting or crouching down occ
asionally. The
second phase is critical phase that very short
period and most important of fall event.
This
phase can be detected by the body movement
toward on the ground or by the impact shock
with the floor.
Next phase is the post fall phase
whichca
n be detected by motionless of human
body after lying on the ground, lying position
or by a motion absence significantly.
phase is the
recovery phase that faller get up
from falling dawn.
Explicitly, fall event in
human can be described as the
rapid movement
from standing or sitting to lying position.
During the critical fall phase there is free fall
temporal time that continuous vertical speed
increasingly. Wu [30] had
showed that vertical
and horizontal speeds are three times higher
during a
fall than other controlled movement
and both speeds will increase near
simultaneously during a fall.
3. Classification of fall detection
techniques
Fig. 2 Classif
ication of fall detection
techniques [14]
As Fig. 2, Mubashir et al. [14]
different types of fall that help to
understanding of existing approaches and lead
to new algorithm designation
.
detection methods can be explained and
categorized into three different kinds as a
hierarchy of fall detection m
ethods.
device based, ambience sensor based and
computer vision based are the three categories
of fall detection methods.
Firstly
, ambience devices can be classified into
presence (vibrations) and posture (audio and
video) based sensors. For
example, v
and sound based fall detector can be used to
detect falls but depends on the vibration of
floor.
They obtained high detection rates, but
phase can be detected by the body movement
toward on the ground or by the impact shock
Next phase is the post fall phase
n be detected by motionless of human
body after lying on the ground, lying position
or by a motion absence significantly.
Last
recovery phase that faller get up
Explicitly, fall event in
rapid movement
from standing or sitting to lying position.
During the critical fall phase there is free fall
temporal time that continuous vertical speed
showed that vertical
and horizontal speeds are three times higher
fall than other controlled movement
and both speeds will increase near
3. Classification of fall detection
ication of fall detection
had specified
different types of fall that help to
understanding of existing approaches and lead
.
Existing fall
detection methods can be explained and
categorized into three different kinds as a
ethods.
Wearable
device based, ambience sensor based and
computer vision based are the three categories
, ambience devices can be classified into
presence (vibrations) and posture (audio and
example, v
ibration
and sound based fall detector can be used to
detect falls but depends on the vibration of
They obtained high detection rates, but
may not detect low impact of real falls.
Moreover
, fall person can call helper using a
push butto
n but it is useless if the person is
insensible or immobilized after the fall
Second
ly, wearable sensors can be classified
into posture based and motion (inactivity)
based devices.
For example, automatic
wearable devices are more interesting becaus
it’s not need intervention by human.
them are based on accelerometers which detect
the magnitude and the direction of the
acceleration or based on gyroscopes which
track the body movement.
kinds of sensors are often embarrassing
wear, and require batteries which need to be
replaced or recharged frequently for adequate
functioning, that are important hindrance of
these technologies. Last
ly, the vision based
devices can be classified into shape change,
posture, 3D head motion,
spatiotemporal.
This kind of devices
techniques and encouraging solution for fall
detection, as no need body
inexpensive cameras example webcam, the
video sequences will contain a high video
compression which c
an generate artifacts or
noises in the image.
4. Visual based tracking systems
According to vision based device mentioned
above, there are several techniques that are
useful for detecting falling down of elderly
people.
As the following details, we have
reviewed and compared different human
motion tracking based on computer vision for
understanding of fall detection techniques.
Fig. 3
Classification of human motion tracking
[31]
may not detect low impact of real falls.
, fall person can call helper using a
n but it is useless if the person is
insensible or immobilized after the fall
[20].
ly, wearable sensors can be classified
into posture based and motion (inactivity)
For example, automatic
wearable devices are more interesting becaus
e
it’s not need intervention by human.
Most of
them are based on accelerometers which detect
the magnitude and the direction of the
acceleration or based on gyroscopes which
track the body movement.
However, these
kinds of sensors are often embarrassing
to
wear, and require batteries which need to be
replaced or recharged frequently for adequate
functioning, that are important hindrance of
ly, the vision based
devices can be classified into shape change,
posture, 3D head motion,
Inactivity, and
This kind of devices
provides
techniques and encouraging solution for fall
detection, as no need body
-worn devices. With
inexpensive cameras example webcam, the
video sequences will contain a high video
an generate artifacts or
4. Visual based tracking systems
According to vision based device mentioned
above, there are several techniques that are
useful for detecting falling down of elderly
As the following details, we have
reviewed and compared different human
motion tracking based on computer vision for
understanding of fall detection techniques.
Classification of human motion tracking
Human motion tracking using sensor
technologies had been classified as in Fig.3 by
Zhou and Hu [31]. In general, a tracking
system can be non-visual, visual based or a
combination of both. Computer vision based
sensors are commonly applied to enhance
precision in position estimation. Visual
tracking systems can be classified as either
visual marker based or marker-free visual
based, depending on whether or not indicators
need to be attached to body parts, as following
details provide a brief description of them.
First system, Visual marker based tracking is a
technique where cameras are adapted to track
human movements, with identifiers position
upon the human body. As the human skeleton
is a highly jointed structure, twists and
rotations generate movement at high degrees-
of-freedom. Consequently, each body part
controls an unpredictable and complex motion
trajectory, which may lead to contradictory and
unreliable motion estimation. Moreover,
disordered scenes or various lighting most
likely divert visual attention from the real
position of a marker. Visual marker based
tracking is preferred to these kind of conditions
that are frequently used as a human motion
analysis standard due to their correct position
information. One weakness of marker based
using optical sensors is unable to detect joint
rotation or body part overlapping and unable to
render 3-D [22].
Second system, marker-free visual based
tracking systems utilize optical sensors
(camera) that can produce continuous high
resolution image to detect movements of the
human body. Currently cameras are popularly
used in surveillance applications with low cost
and flexible to configure by user. However this
technique requires complicated computation to
error reduction, 3-D rendering, and decreasing
of data latency [4]. As common cameras
provide inadequate bandwidth accurate data
representation, high speed cameras are
required [2].
Lastly, Combination tracking systems take
advantage of marker based and marker-free
based technologies to reduce errors coming
from using individual technology [24].
5. Related works on fall detection
techniques based on computer vision
Generally, several kinds of techniques were
used to detect falling down event in pre-fall,
critical-fall or post fall phases as example
Bayesian, Centroidand Fuzzy. Some of them
can only been used in one phase but some were
used in several phases. Following details are
briefly example of techniques that were used in
fall detection system.
5.1 Bayesian Technique
Cucchiara et al. [5] proposed a human posture
classifier using Bayesian technique to detect
falling in critical through post-fall phases. The
technique starts from a visual object extracted
by low level tasks and classified as a person,
models human posture as probabilistic
projection maps (PPMs). The classification is
finished by using a Bayesian framework based
on a dimension of the resemblance between the
current projection histograms of the silhouette
of the analyzed visual objects and the PPMs.
The visual object based classification is more
analyzed by taking into address the integration
along time that used a state-transition graph
and a confidence measure.
Moreover, Rimminen et al. [17]also proposed a
method using a floor sensor based on near-
field imaging to detect fall event in critical
through post-fall phases. The technique uses
shape, size, and magnitude of the patterns for
classification. The person poses were estimated
using Bayesian filtering instead of using direct
feature classification. The state evolution was
modeled as a two-state Markov chain and
recursively computes the probability
distribution of the current state, given all the
features observed so far, that were estimated
from a training set.
5.2 Centroid
Fu et al.[8] proposed an address-event
temporal contrast fall detection and using this
detector to detect falls in critical fall phase.
They use a temporal moderate of the motion
events from the asynchronous temporal
contrast (ATC) vision sensor, here referred as
centroid event, to track fall risks and evaluate
its dynamics. Centroids are an effective way to
measure object motion in space and can be
computed as temporal averages of a series of
events. Currently, the modern 3D range
imaging (RIM) , integrates distance
measurement as well as imaging aspects,
using time-of-flight (TOF) cameras has
become more popular as an appropriate
choice in monitoring applications.
Furthermore, Leone et al. [9]had presented a
preceding development and testing of a
real-time system for the detection of falls,
based on a self calibrated 3D TOF camera
that does not require calibration objects,
landmarks or user intervention and is easily
install. A fall event was detected when the
distance of the silhouette's centroid with
aspect to the floor plane. The proposed
method for fall detection worked when a
whole human silhouette was detected, and a
partial occlusion happened. That means this
technique also detected fall event in critical fall
phase through post-fall phase.
Moreover, a feature-based compressed-domain
fall-down detection method for intelligent
surveillance applications had been presented
by Lin and Ling [10]. The method implies two
steps, compressed-domain object extraction
and fall incident detection in post-fall phase.
Three feature values, the change ratio of the
centroid of a human object, the change ratio of
the maximum of vertical projection histogram,
and the duration of an event detected, were
used to analyze and locate fall-down events,
after extracting moving human objects.
5.3 Eigen-space
Software based upon a spatio-temporal motion
representation, Motion Vector Flow Instance
(MVFI) templates was proposed by Olivieri et
al. [16] that capture relevant velocity
information by extracting the dense optical
flow from human action video sequences.
Automatic recognition is achieved by firstly
projecting each human action video sequence,
secondly including of approximately 100
images, thirdly into an accepted eigen-space,
and lastly performing supervised learning to
train multiple actions from a large video
database. They showed both of canonical
conversion with principal component analysis
(PCA) and linear discriminant analysis (LDA)
of image sequences provides excellent action
differentiation. Additionally, they also
expressed that by including both the magnitude
and direction of the velocity in critical fall
phase into the motion vector flow instances
(MVFI), sequences with immediate velocities.
An efficient method for activity recognition
based on fusion of integrated time motion
images and eigen-space technique, mainly
dedicated to fall detection was proposed by
Foroughi et al. [7]. Integrated Time Motion
Image (ITMI), includes motion and time of
motion happening, is a type of spatio-temporal
database. Applying eigen-space technique to
ITMIs leads in extracting eigen-motion and
MLP Neural Network is used for accurate
classification of motions and definition of a
fall event in critical fall phase. The integration
of motion and eigen-space technique, gives
particular information of human activities.
5.4 Fuzzy Logic
Anderson et al. [1] proposed the results in
human understandable information and
reliabilityof activities for the purpose of
monitoring the “well-being” of elderly people.
They used silhouettes from multiple cameras to
build a 3-D estimation of the human by
extracting feature from voxelperson and used
together with fuzzy inference to specify the
state of fall in critical through post-fall phases.
The resulting fuzzy rule base outputs are used
to generate temporal linguistic summarizations
and then temporally processed.
5.5 Gaussian Mixtures
A model was presented by McKenna and
Charif [12] for automatically learning a
context-specific spatial model in terms of
semantic regions, particularly inactivity zones
and entry zones. Maximum a posteriori
assessment of Gaussian mixtures is used in
fusion with minimum feature length for
selection of the number of mixture
components. Learning is functioned using
expectation-maximization algorithms to
maximize inflicted possibility functions. That
consolidates prior knowledge of the size and
shape of the semantic areas and facilitates a
one-to-one compatibility between the Gaussian
mixture components and the areas. Results of
contextual model enables human-readable
summaries of activity are presented using
overhead camera sequences tracked using a
particle filter that can be produced and unusual
inactivity in post-fall phase to be detected.
Rougier et al. [19] also presented a GMM
classification method to detect falls in critical
through post-fall phases by analyzing human
shape contortion during a video sequence. The
edge point matching step of shape context is
determined to contortion and other
segmentation difficulties. A useful tool to
detect falls is human shape contortion because
they can be sensitive to lacking matching
points only credible matching points were kept
for shape contortion assessment. The silhouette
of human along the video sequence was
tracked by using a shape matching technique.
The peak of fall is an important feature to
indicate a fall in critical phase, but the lack of
significant movement after the fall is also
important for robustness when obstructions
occur. Finally, a Gaussian mixture model was
detected from common activities of falls.
5.6 Hidden Markov Model
Hierarchical Hidden Markov Model (HHMM)
was proposed by Thome and Miguet [25]
whose first layer states are related to the
orientation of the tracked person in critical and
post fall phase. The heart of model
contribution is exploring a stable method for
robustly connecting the observation vector to
the human poses, and carefully studies the
relationship between angles in the 3D world
and their projection onto the image plane.
Several efficiencies of the algorithm were
resulted by pointing out its ability to correctly
recognize a falling down person from another
states, and ability to run in an unspecified
configuration. Not only HHMM but they also
proposed a multi-view methodto achieve
automatic detection of a falling person in video
sequences, where motion is modeled using a
layered hidden Markov model (LHMM) [26].
The posture classification is accomplished by a
fusion unit that integrates the decision
provided by the independently processing
cameras in a fuzzy logic context. The method
is optimized in each view, given plane by
performing a metric image validation, making
it able to easily separate and robust features,
and being suitable for real-time purpose.
5.7 Neural Network Technique
Bromiley et al. [3] proposed the development
of a vision system to detect natural events in a
low-resolution image stream that concerned
the assessment of algorithmic design decisions
to increase detection reliability. As the
approach of calculating vertical velocities from
the infrared images and using a neural network
to classify fall or non-fall in critical through
post fall phase is sufficient to produce an
actually fall detector. Close experimentation of
the scenarios showed that most were due to
movements of specific simulations, to
movements immediately happening after a fall,
or to changes in the viewing angle of the
cameras during the simulations. A comparison
of the basic physics of falls to the ground that
event can be explained in terms of the
distributions of vertical velocities they
generate, and the performance of the nearest
neighbor classifier, an honest classifier.
5.8 Posture Based Technique
A multi camera vision system for detecting and
tracking human and recognizing risk behaviors
and events in critical and post fall phases had
been presented by Cucchiara et al. [6]. As the
new hardware technologies and in particular
digital cameras are now affordable and
popularly used as tools for automatically
assuring the human safety. Warping human
silhouette technique is proposed to exchange
visual information between partially
overlapped cameras whenever a camera
handover occurs.
Liu and Zuo[11] also proposed an improved
algorithm of automatic fall detection that use
human aspect ratio, effective area ratio and
center variation rate, three features which can
effectively prevent misjudgments and greatly
increase the accuracy of detection results. This
algorithm has a less computing complexity and
is easy to implement and good robustness. The
detection system that uses a MapCam (omni-
camera) to capture images and performs image
processing over the images was proposed by
Miaou et al. [13]. The system is used to
simultaneously capture 360 degree scene,
remove any blind viewing zone, and use
personal information to enhance the
recognition rate. As the personal information
(height, weight, and electronic health history),
researchers can adjust the detection sensitivity
on a case by case basis to reduce errors, and
put more attention on the elderly with special
diseases or conditions. Additionally, Rougier
et al. [18] also proposed a method to detect
falls is based on a combination of motion
history and human shape variation that
algorithm provides promising results on video
sequences of daily activities and simulated
falls.
Tao et al. [23] presented an intelligent video
surveillance system to detect human fall
incidents as encloses of a vision component
which can reliably detect and track moving
human in camera view, and an event-inference
module which extract observation sequences of
human features for possible falling behavioral
signs. The aspect ratio of human as observation
feature, based on which fall incidents are
detected as immediately changes in the feature
space was extracted.
5.9 Support Vector Machine: SVM
A MEMS-based human airbag system was
presented by Shi et al. [21] that using Micro
Inertial Measurement Unit (IMU) for the
detection of complicated human motions and
the recognition of a falling down motion in
critical phase, which can be used to release an
airbags. The system records human motion
information and the analysis of falls using a
high-speed camera, a lateral fall can be
determined by gyro threshold. Researchers set
up a human motion database that includes falls
and other normal motions and use a SVM
training process to classify falls and other
normal motions with a SVM filter that is
developed by an embedded digital signal
processing (DSP) system for real-time fall.
Moreover, Williams et al. [29] proposed a
network of overlapping smart cameras that
uses a decentralized process for computing
inter-image homographies and allows to
reporting fall location in 2D world coordinates
by calibrating only one camera. Even some of
the more complicated fall detection techniques
tend to one particular feature of a human being
monitored that feature is their aspect ratio, or
the width of body posture divided by height. It
is easily extracted from foreground
segmentation, and is a good method of
indicating whether a human is upright or in a
more horizontal position with a simple
threshold.
5.10 Velocity
Wu [30] presented the study of identify unique
features of the velocity profile during normal
and abnormal activities so as to make the
automatic detection of falls during the critical
phase of a fall happen. The horizontal and
vertical velocities at various areas of the trunk
were measured. Additionally, researchers [11]
found that the horizontal velocity and vertical
velocity of the trunk during normal activities
were within a well-controlled distances, and
that when the velocity in one direction
increased, the velocity in the other direction
normally did not. These two velocity
characteristics, that is, the dimension change
and the timing of the dimension change of both
horizontal and vertical velocities, could be
used to differentiate fall movements from
normal activities during the critical phase of
the fall.
6. Summarization of phase detection
As Table 1, summarization of phase detection
using several techniqueshas been presented.
There are 18 items of critical phase detection
and 15 items of post-fall phase detection which
several techniques. Most of them detect falling
in critical and post fall phasessequentially that
means pre-fall and early critical fall phase
should be considered.
Table 1: Classification of Phase Detection
Detection
Techniques Critical
Phase Post-fall
Phase
Bayesian [5, 17] [5, 17]
Centroid [8, 9] [9, 10]
Eigen-space [7, 16]
Fuzzy [1] [1]
GMM [19] [12, 19]
HMM [25, 26] [25, 26]
Neural
Network [3] [3]
Posture [6, 13, 18,
23] [6, 13, 18,
23]
SVM [21] [29]
Velocity [11, 30]
Total 18 15
7. Conclusion and future work
We have reviewed different fall detection
techniques as Bayesian, Centroid, Context
Aware, Eigen-space, Fuzzy Logic, GMM,
HMM, Neural Network, Posture, SVM, and
Velocity. Several fall detection techniques are
good solution for fall detection with high rates
but mostly detect on lately critical phase or
post-fall phase that mean fall person had been
injured already. To reduce the risk of falling,
the early detection of fall in pre-fall or critical
fall phase therefore advances the interest of
researchers. Multi-camera systems could also
be used to improve the recognition results by
combining information from several cameras
to take a decision. Some new sensor
technologies could also be used to improve the
recognition results in pre fall and early critical
fall phase by combining information from
several cameras and sensors like PrimeSence,
Kinect or Xtion sensors that can provide color,
depth, audio and video stream. Not only
standard color camera but a depth camera also
can send a stream integrated of the distance
between the camera plane and the nearest
object found. Moreover, it features a
microphone array that is possible to determine
the direction of the audio source and optimizes
skeleton tracking to recognize users as six
people can be detected and one or two human
can be tracked at one time.
References
[1] Anderson, D., Luke, R. H., Keller, J. M.,
Skubic, M., Rantz, M., &Aud, M. (2009).
Linguistic summarization of video for fall
detection using voxel person and fuzzy
logic. Computer Vision and Image
Understanding, 113(1), 80-89.
[2] Bhatnagar, D. K. (1993). Position trackers
for Head Mounted Display systems: A
survey. University of North Carolina,
Chapel Hill TR93, 10.
[3] Bromiley, P. A., Courtney, P., & Thacker,
N. A. (2002). Design of a visual system for
detecting natural events by the use of an
independent visual estimate: a human fall
detector. Empirical Evaluation Methods in
Computer Vision, 50.
[4] Bryson, S. (1996). Virtual reality in
scientific visualization. Communications of
the ACM, 39(5), 62-71.
[5] Cucchiara, R., Grana, C., Prati, A., &
Vezzani, R. (2005). Probabilistic posture
classification for human-behavior analysis.
Systems, Man and Cybernetics, Part A:
Systems and Humans, IEEE Transactions
on, 35(1), 42-54.
[6] Cucchiara, R., Prati, A., & Vezzani, R.
(2007). A multi
camera vision system for
fall detection and alarm generation. Expert
Systems, 24(5), 334-345.
[7] Foroughi, H., Naseri, A., Saberi, A., &
Yazdi, H. S. (2008, October). An
eigenspace-based approach for human fall
detection using integrated time motion
image and neural network.In Signal
Processing, 2008. ICSP 2008. 9th
International Conference on (pp. 1499-
1503). IEEE.
[8] Fu, Z., Culurciello, E., Lichtsteiner, P., &
Delbruck, T. (2008, May). Fall detection
using an address-event temporal contrast
vision sensor. In Circuits and Systems,
2008. ISCAS 2008. IEEE International
Symposium on (pp. 424-427). IEEE.
[9] Leone, A., Diraco, G., & Siciliano, P.
(2011). Detecting falls with 3D range
camera in ambient assisted living
applications: A preliminary study. Medical
engineering & physics, 33(6), 770-781.
[10] Lin, C. W., & Ling, Z. H. (2007, August).
Automatic fall incident detection in
compressed video for intelligent homecare.
In Computer Communications and
Networks, 2007. ICCCN 2007. Proceedings
of 16th International Conference on (pp.
1172-1177). IEEE.
[11] Liu, H., &Zuo, C. (2012). An Improved
Algorithm of Automatic Fall Detection.
AASRI Procedia, 1, 353-358.
[12] McKenna, S. J., &Charif, H. N. (2004).
Summarising contextual activity and
detecting unusual inactivity in a supportive
home environment. Pattern Analysis and
Applications, 7(4), 386-401.
[13] Miaou, S. G., Sung, P. H., & Huang, C. Y.
(2006, April). A customized human fall
detection system using omni-camera images
and personal information. In Distributed
Diagnosis and Home Healthcare, 2006.
D2H2. 1st Transdisciplinary Conference on
(pp. 39-42). IEEE.
[14] Mubashir, M., Shao, L., & Seed, L.
(2013). A survey on fall detection
Principles and approaches. Neurocompu-
ting, 100, 144-152.
[15] Noury, N., Rumeau, P., Bourke, A. K.,
ÓLaighin, G., & Lundy, J. E. (2008). A
proposal for the classification and
evaluation of fall detectors. Irbm, 29(6),
340-349.
[16] Olivieri, D. N., Gómez Conde, I., & Vila
Sobrino, X. A. (2012). Eigenspace-based
fall detection and activity recognition from
motion templates and machine learning.
Expert Systems with Applications, 39(5),
5935-5945.
[17] Rimminen, H., Lindstrom, J., Linnavuo,
M., &Sepponen, R. (2010). Detection of
falls among the elderly by a floor sensor
using the electric near field. Information
Technology in Biomedicine, IEEE
Transactions on, 14(6), 1475-1476.
[18] Rougier, C., Meunier, J., St-Arnaud, A., &
Rousseau, J. (2007, May). Fall detection
from human shape and motion history using
video surveillance. In Advanced
Information Networking and Applications
Workshops, 2007, AINAW'07. 21st
International Conference on (Vol. 2, pp.
875-880).IEEE.
[19] Rougier, C., Meunier, J., St-Arnaud, A., &
Rousseau, J. (2011). Robust video
surveillance for fall detection based on
human shape deformation. Circuits and
Systems for Video Technology, IEEE
Transactions on, 21(5), 611-622.
[20] Rougier, C., St-Arnaud, A., Rousseau, J.,
&Meunier, J. (2009). Video Surveillance
for Fall Detection.
[21] Shi, G., Chan, C. S., Li, W. J., Leung, K.
S., Zou, Y., & Jin, Y. (2009). Mobile
human airbag system for fall protection
using MEMS sensors and embedded SVM
classifier. Sensors Journal, IEEE, 9(5), 495-
503.
[22]Sturman, D. J., & Zeltzer, D. (1994). A
survey of glove-based input. Computer
Graphics and Applications, IEEE, 14(1),
30-39.
[23] Tao, J., Turjo, M., Wong, M. F., Wang,
M., & Tan, Y. P. (2005, December). Fall
incidents detection for intelligent video
surveillance. In Information, Communi-
cations and Signal Processing, 2005 Fifth
International Conference on (pp. 1590-
1594). IEEE.
[24] Tao, Y., & Hu, H. (2003, September).
Building a visual tracking system for home-
based rehabilitation. In Proc. of the 9th
Chinese Automation and Computing
Society Conf.In the UK.
[25] Thome, N., &Miguet, S. (2006,
December). A HHMM-based approach for
robust fall detection.In Control,
Automation, Robotics and Vision, 2006.
ICARCV'06. 9th International Conference
on (pp. 1-8). IEEE.
[26] Thome, N., Miguet, S., &Ambellouis, S.
(2008). A real-time, multiview fall detection
system: A LHMM-based approach. Circuits
and Systems for Video Technology, IEEE
Transactions on, 18(11), 1522-1532.
[27] WHO (2012). Good health adds life to
years: Global brief for World Health Day
2012. retrived form
(http://www.who.int/ageing/publications/w
hd2012_global_brief/en/)
[28] WHO at Geneva (2002). retrived form
(http://www.who.int/ageing/active_ageing/e
n/index.html)
[29] Williams, A., Ganesan, D., & Hanson, A.
(2007, September). Aging in place: fall
detection and localization in a distributed
smart camera network. In Proceedings of
the 15th international conference on
Multimedia (pp. 892-901).ACM.
[30] Wu, G. (2000). Distinguishing fall
activities from normal activities by velocity
characteristics. Journal of biomechanics,
33(11), 1497-1500.
[31] Zhou, H., & Hu, H. (2008). Human
motion tracking for rehabilitation-A survey.
Biomedical Signal Processing and Control,
3(1), 1-18.
... Human gesture detection is one of the considerably challenging issues in the pre-impact fall detection system. According to [1] and [2,3], the second foremost cause of incidental death comprehensive is unintended slips [4], which is a critical cause of individual damage, especially for the esteemed. Consequently, many investigations in healthcare are improving the pre-impact fall detection system to protect those who are possibly concerned. ...
... fall detection approach: The pre-impact fall detection technique must confound many complications to improve a practical approach [10]. Some particular problems are occlusion, obtrusion, and overlapping in the vision-based method [2]. Further associated troubles are privacy interests, price, noise, calculation complexity, and definition of the threshold-based values [10]. ...
... The Euclidean metric has been used to calculate straight-line distance between different feature skeletal joints. The DL and DR range among left and right ankle joint position as equation number (1), and DRJ and DLJ is the range on both sides of ankles and spine position as defined in equation number (2). Next, the circle region in figure 1 means for the BSA; meanwhile, COG is moving inside the BSA, people are safe, and vice versa, who is possibly unsecured or unstable while COG is running outside the BSA. ...
Article
Full-text available
The problem of falling among the elderly has significantly increased continuously. Researchers are trying to find a suitable way to monitor and alert before a fall to protect the body using an airbag, for example, to prevent the consequences of falls. In this paper, pre-fall surveillance and detection using three Kinnects installed at different angles enable detection without body tracking devices, which causes trouble. In addition, Kinnect detection uses only the head position, the center of gravity, and the position of the feet to calculate the base support area, so there are no privacy violations. The model used to predict the fall event was a logistic regression analysis that used predictive variables of head displacement versus dynamic threshold and body center of gravity versus base support area. From the pre-impact fall predictor experimental 31 results, the accuracy was 98.17%, the sensitivity was 87.97%, and the specificity was 98.98%. Therefore, it can be concluded that the developed system can detect pre-impact fall events using the logistic regression model and can function at the specified time according to the objective.
... Human motion detection is one of the most challenging topics in the fall detection system. According to Mastorakis & Makris [11] and Otanasap & Boonbrahm [13], the second leading cause of accidental death extensive is unintentional falls [18], which is a vital cause of personal harm, particularly with the venerable. Accordingly, many studies in healthcare are achieving on the improvement of the pre-fall detection system to secure the protection of those who are possible to be concerned. ...
... 2.1 Pre-impact fall detection system: Preimpact fall detection system has to overcome many difficulties to improve an effective system [15]. Some of the particular difficulties are obtrusion, occlusion, and overlap in the vision-based method [13]. Additional associated problems are interests in privacy, price, noise, computation complexity, and definition of the threshold based values [15]. ...
Article
Otanasap, N. (2019). Pre-Impact Fall Detection System Using Real Time Dynamic Threshold and Human Body Bounding Box. SAU JOURNAL OF SCIENCE & TECHNOLOGY, 5(1), 49-61. Retrieved from https://www.tci-thaijo.org/index.php/saujournalst/article/view/170324 This work contributes on the fusion of multiple Kinect based skeletons, based on dynamic threshold and bounding box posture analysis which is the only research work reported so far. As the second leading cause of accidental death extensive is unintentional falls, which is a vital cause of personal harm, particularly with the venerable. Accordingly, many studies in healthcare are achieving on the improvement of the pre-fall detection system to secure the protection of those who are possible to be concerned. Pre-impact fall detection system has to overcome many difficulties to improve an effective system. Some of the particular difficulties are obtrusion, occlusion, and overlap in the vision-based method. In this research, the purpose of adopting the bounding box and head velocity method compare with a real-time dynamic threshold is for analyzing the fall and non-fall incident accurately. Furthermore, the skeleton joint position provided by multiple Kinect viewpoints are utilized for the reason of resolving in obtrusion, occlusion, and overlap issues without demanding of markers. Though, the various fuzzy rule base methods also are applied for the final decision of lead time detection and triggering fall alarm. The demonstration of subjects completion is performed 1,100 actions were included 700 times for activities of daily living and 400 times for falling. All activities performed by ten different volunteers, seven healthy young males, and three healthy young females. The results have shown that 98.55% of the proposed method is higher accurately detected. However, the proposed method provided the lowest specificity at 97.71%, vice versa it offered the highest sensitivity at 100.00%. It implies that during system provided higher accuracy and sensitivity in pre-impact fall detection, the recognized precision of normal activities will be reduced. Moreover, the multiple Kinect methods not only provide higher accuracy and sensitivity but also offer higher average lead-time as 505.86 ms.
... But the main problem [11] of these kinds of technology is the fall victims have forgotten to wear the devices while the accidents occur. Therefore, some researchers attempt to solve these problems by using computer vision techniques such as omni camera [5] and Kinect camera [6,8,10] that do not require the person to wear any devices. In particular, Kinect cameras for human posture classification and fall detection are becoming popular as the cameras are relatively inexpensive and free Software Development Kit (SDK) are also available for development purposes. ...
Conference Paper
Full-text available
Falling accidents, including slipping, tripping, and falling, are the primary reason for injury related to death not only for the elderly but for young people or worker happening at the workplace also. If a falling accident can be early detected in pre-fall or critical fall phase, called pre-impact fall detection, it will be beneficial such as conducting airbag inflation. Furthermore, various detection methods, with an uncomplicated threshold detection method, do maximize the valid positive prediction values but the lead-time, time before subject impacts to the floor, will likely increase the chance of false alarms. Consequently, the researcher found that the usage of the adaptive threshold may reduce false alarms. In this paper, the dynamic threshold method, automatically adjustable threshold for pre-impact fall detection in a wearable device, has been proposed and experimented. For our evaluation, 192 instances of several kinds of activity of daily living and falling, were captured. All activities were performed by 6 different young healthy volunteers, 4 males and 2 females, aged between 19 and 21. The several experiments were conducted for performance evaluation, including sensitivity, specificity, and accuracy measurements. The results of the proposed method can detect the pre-impact fall from normal activities of daily living with 99.48% sensitivity, 95.31% specificity, and 97.40% accuracy with 365.12 msec of lead time. The results confirm that our proposed method with an automatically adjustable threshold based on motion history is suitable for using in pre-impact fall detection system than a fixed threshold based method.
... But the main problem [11] of these kinds of technology is the fall victims have forgotten to wear the devices while the accidents occur. Therefore, some researchers attempt to solve these problems by using computer vision techniques such as omni camera [5] and Kinect camera [6,8,10] that do not require the person to wear any devices. In particular, Kinect cameras for human posture classification and fall detection are becoming popular as the cameras are relatively inexpensive and free Software Development Kit (SDK) are also available for development purposes. ...
Conference Paper
Full-text available
Fall prevention and detection system have to subjugate many challenges to develop an efficient those systems. Some of the severe problems are obtrusion, occlusion, and overlay in a vision based system. Other associated issues are privacy, cost, noise, computation complexity, and definition of threshold values. Estimating human motion using vision-based usually involves with partial overlay, caused either by the direction of viewpoint between objects or body parts and camera, and these issues have to be taken into consideration. This paper proposes the use of dynamic threshold based and bounding box posture analysis method with multiple Kinect cameras setting for human posture analysis and fall detection. The proposed work only uses two Kinect cameras for acquiring distributed values and differentiating activities between normal and falls. If the peak value of head velocity is higher than the dynamic threshold value, bounding box posture analysis will be used to confirm fall occurrence. Furthermore, information captured by multiple Kinect placed in right angle will address the skeleton overlay problem due to single Kinect. This work contributes to the fusion of multiple Kinect based skeletons, based on dynamic threshold and bounding box posture analysis, which is the only research work reported so far.
Conference Paper
One of the primary reasons for injury related to death not only in elderlies but also in young people is slip trip and fall accidents. It will be advantageous if fall accidents can be detected in pre-fall and critical fall phase long enough before the human body impact to the floor. Pre-impact fall detection method and lead time before effects are critical factors that have been used to save a person who takes risks. In this paper, multiple viewpoints in vision-based sensing that provided by multiple Kinect© sensors using dynamic threshold based and center of gravity are proposed. The acceleration of head position will trigger Pre-fall detection alert compared with a dynamic threshold based approach and range of center of gravity compared with a base of the support area. Moreover using the blinding technique without video stream recording is one of useful feature in our vision-based approach for reducing privacy issue. Not only fall actions but also a series of normal activities in daily living such as sitting, bending and laying were performed by 10 young adult volunteers. Results from the experiments indicate that the proposed method lead time is about 500 milliseconds, which is faster than the previous proposition, and it is appropriate for airbags inflation triggering.
Conference Paper
As falling is the most important issue that faces elderly people all over the world, this paper proposes a detection system for falling based on Machine Learning (ML). In the proposed system, a dataset of videos containing falling actions has been utilized via dividing each video into many shots that are consequently being converted into gray-level images. Then, for detecting the moving objects in videos, the foreground is firstly detected, then noise and shadow are deleted to detect the moving object. Finally, a number of features, including aspect ratio and falling angle, are extracted and a number of classifiers are being applied in order to detect the occurrence of falling. Experimental results, using 10-fold cross validation, shown that the proposed falling detection approach based on Linear Discriminant Analysis (LDA) classification algorithm has outperformed both support vector machines (SVMs) and Knearest neighbor (KNN) classification algorithms via achieving falling detection with accuracy of 96.59 %.
Conference Paper
Full-text available
More than ten millions elderly people fall each year. Falls are the important cause of injury related to death and set of symptoms. Most of "fall detection" systems are focused on critical and post-fall phase which mean that the faller may already be injured. In this study, we propose an early fall detection in critical fall phase using velocity characteristics, collected by Kinect sensor with 30 frames per second. A series of normal and falling activities were performed by 5 volunteers in first experiment and 11 volunteers in second experiment. The fall velocity based point was calculated by the first experiment as 50 postures, 2210 frames recorded. The result from the first experiment, velocity ratio 90.35 pixels per millisecond was calculated by adding μ with σ, was defined as velocity fall detection based point for the second experiment. In the second experiment, the fall activities were detected at 85.07 percent from 134 fall activities. The mean time of fall activities that was detected before dash to the ground is 391.15 milliseconds. The detected mean time may be useful to developing a preventive fall system to protect the faller before injured i.e. wearable airbag system. For high accuracy prediction, automatically adjusted vertical fall based point by train dataset of individual person will be investigated in future work.
Article
Full-text available
Fall detection is an important problem in the application research of video surveillance. Quick and effective ways of getting the information of the people falling to the ground can help the people get timely assistance to reduce further injury. This paper proposes an improved algorithm of automatic fall detection. Three features — human aspect ratio, effective area ratio and center variation rate are used in the proposed algorithm, which can effectively prevent misjudgments and greatly increase the accuracy of detection results. This algorithm has a less computing complexity and is easy to implement. Experimental results show that the algorithm has good robustness.
Article
Full-text available
We describe the development of a vision system to detect natural events in a low-resolution image stream. The work involved the assessment of algorithmic design decisions to maximise detection re- liability. This assessment was carried out by comparing measures and estimates made by the system under development with measures obtained independently. We show that even when these independent measures are themselves noisy, their independence can serve to guide rational design decisions and allow performance estimates to be made. Although presented here for one particular system design, we believe that such an approach will be applicable to other situations when an image-based system is to be used in the analysis of natural scenes, in the absence of a precise ground truth. The detector operated by recognising the patterns of vertical velocities present during a fall by a human subject, and issuing a fall detection warning when such a pattern was detected in the output from the thermal sensor. The initial identification and the detection of these patterns was performed automatically using a multi-layer perceptron (MLP) neural network, thus removing any requirement for in-depth study of the dynamics of human movement. Image data showing simulated fall and non-fall (e.g. sitting) scenarios were captured from both the infrared sensor and a colour CCD camera. Vertical velocity measurements were extracted from both sets of images, and the colour data were used as a gold standard to demonstrate a correlation between the infrared velocity estimates and the physical velocities present in the scene. Despite the fact that the velocity measurements extracted from the colour data did not represent a genuine ground truth, we show that this form of performance evaluation can be used to guide algorithmic development in an objective manner. Finally, a range of neural networks were trained on a sub-set of the infrared data and tested on the remainder, and the best-performing network was identified using receiver-operator curves (ROC). We demonstrate that, using appropriate definitions of true and false detections, it is possible to evaluate the performance of a system for identifying events in a temporal data stream in this manner.
Conference Paper
Full-text available
This paper presents a compressed-domain fall incident detection scheme for intelligent homecare applications. First, a compressed-domain object segmentation scheme is performed to extract moving objects based on global motion estimation and local motion clustering. After detecting the moving objects, three compressed-domain features of each object are then extracted for identifying and locating fall incidents. The proposed system can differentiate fall-down from squatting by taking into account the event duration. Our experiments show that the proposed method can correctly detect fall incidents in real time.
Conference Paper
Full-text available
Nowadays, Western countries have to face the growing population of seniors. New technologies can help people stay at home by providing a secure environment and improving their quality of life. The use of computer vision systems offers a new promising solution to analyze people behavior and detect some unusual events. In this paper, we propose a new method to detect falls, which are one of the greatest risk for seniors living alone. Our approach is based on a combination of motion history and human shape variation. Our algorithm provides promising results on video sequences of daily activities and simulated falls.
Article
Automatic recognition of anomalous human activities and falls in an indoor setting from video sequences could be an enabling technology for low-cost, home-based health care systems. Detection systems based upon intelligent computer vision software can greatly reduce the costs and inconveniences associated with sensor based systems. In this paper, we propose such a software based upon a spatio-temporal motion representation, called Motion Vector Flow Instance (MVFI) templates, that capture relevant velocity information by extracting the dense optical flow from video sequences of human actions. Automatic recognition is achieved by first projecting each human action video sequence, consisting of approximately 100 images, into a canonical eigenspace, and then performing supervised learning to train multiple actions from a large video database. We show that our representation together with a canonical transformation with PCA and LDA of image sequences provides excellent action discrimination. We also demonstrate that by including both the magnitude and direction of the velocity into the MVFI, sequences with abrupt velocities, such as falls, can be distinguished from other daily human action with both high accuracy and computational efficiency. As an added benefit, we demonstrate that, once trained, our method for detecting falls is robust and we can attain real-time performance.
Article
Fall detection is a major challenge in the public health care domain, especially for the elderly, and reliable surveillance is a necessity to mitigate the effects of falls. The technology and products related to fall detection have always been in high demand within the security and the health-care industries. An effective fall detection system is required to provide urgent support and to significantly reduce the medical care costs associated with falls. In this paper, we give a comprehensive survey of different systems for fall detection and their underlying algorithms. Fall detection approaches are divided into three main categories: wearable device based, ambience device based and vision based. These approaches are summarised and compared with each other and a conclusion is derived with some discussions on possible future work.
Article
Falls affect, each year, tens of million of elderly people throughout the world. It can have immediate lethal consequences but also causes many disabling fractures and dramatic psychological consequences which reduce the independence of the person. Falls in the elderly is thus a major public health problem. The “early” detection of fall consequently raises the interest of searchers, as most of elderly fallers cannot return to a standing position on their own following a fall. It is also an interesting scientific problem because it is an ill-defined process. The goals of this study were to classify various approaches used to detect the fall and to point out the difficulty to compare the results of these studies, as there is currently no common evaluation benchmark.
Article
Visual tracking of human movement has attracted increasing attention recently because of its wide spectrum of applications, including athletic and clinical performance analysis, human-computer interface, surveillance, motion capture for games and animation. There are two main techniques in the visual tracking of the human movement community: marker-based tracking and marker-free tracking. This paper presents a visual tracking system, which exploits both marker-based and marker-free methods, to support the home-based rehabilitation program. It provides a useful and complete tracking system to help the patients' who sustain a stroke to recover and improve their mobility at a home environment.
Conference Paper
We present in this paper an intelligent video surveillance system to detect human fall incidents for enhanced safety in indoor environments. The system consists of two main parts: a vision component which can reliably detect and track moving people in the view of a camera, and an event-inference module which parses observation sequences of people features for possible falling behavioral signs. In particular, we extract the aspect ratio of a person as observation feature, based on which fall incidents are detected as abrupt changes in the feature space. Our experiments show that the proposed approach can robustly detect human falls in real time